def get_layers(self):
return [
dfxp.Conv2d_q(
name='conv1',
bits=self.bits,
ksize=[5, 5, 1, 6],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID'
),
dfxp.Conv2d_q(
name='conv2',
bits=self.bits,
ksize=[5, 5, 6, 16],
strides=[1, 1, 1, 1],
padding='VALID',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID'
),
dfxp.Conv2d_q(
name='conv3',
bits=self.bits,
ksize=[5, 5, 16, 120],
strides=[1, 1, 1, 1],
padding='VALID',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.Flatten_q(120),
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
in_units=120,
units=84,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=84,
units=10,
weight_decay=self.weight_decay,
),
]
def get_layers(self):
return [
dfxp.Conv2d_q(
name='conv',
bits=self.bits,
training=self.training,
ksize=[5, 5, 1, 20],
strides=[1, 1, 1, 1],
padding='VALID',
weight_decay=self.weight_decay,
),
dfxp.BatchNorm_q(
name='batch_normolization',
bits=self.bits,
num_features=20,
training=self.training,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID'),
dfxp.Flatten_q(12 * 12 * 20),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
training=self.training,
in_units=12 * 12 * 20,
units=100,
weight_decay=self.weight_decay,
),
dfxp.Dense_q(
name='dense2',
bits=self.bits,
training=self.training,
in_units=100,
units=10,
weight_decay=self.weight_decay,
)
]
def get_layers(self):
return [
# conv1-1
dfxp.Conv2d_q(
name='conv1-1',
bits=self.bits,
ksize=[3, 3, 3, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# conv1-2
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv1-2',
bits=self.bits,
ksize=[3, 3, 128, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# pool1
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv2-1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv2-1',
bits=self.bits,
ksize=[3, 3, 128, 256],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# conv2-2
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv2-2',
bits=self.bits,
ksize=[3, 3, 256, 256],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# pool2
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv3-1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv3-1',
bits=self.bits,
ksize=[3, 3, 256, 512],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# conv3-2
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv3-2',
bits=self.bits,
ksize=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# pool3
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
dfxp.Flatten_q(512*4*4),
# dense1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
in_units=512*4*4,
units=1024,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# dense2
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense2',
bits=self.bits,
in_units=1024,
units=1024,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# softmax
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=1024,
units=10,
weight_decay=self.weight_decay,
),
]
def get_layers(self):
return [
# conv1
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv1',
bits=self.bits,
ksize=[5, 5, 3, 64],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv2
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv2',
bits=self.bits,
ksize=[5, 5, 64, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv3
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv3',
bits=self.bits,
ksize=[5, 5, 128, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
dfxp.Flatten_q(128*4*4),
# dense1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
in_units=128*4*4,
units=400,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# softmax
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=400,
units=10,
weight_decay=self.weight_decay,
),
]
def __init__(self, model, dataset, dataset_name, logger, params):
self.n_epoch = params.n_epoch
self.exp_path = params.exp_path
self.logger = logger
self.graph = tf.Graph()
with self.graph.as_default():
global_step = tf.train.get_or_create_global_step()
self.lr_scheduler = LearningRateScheduler(params.lr,
params.lr_decay_epoch,
params.lr_decay_factor)
optimizer = tf.train.MomentumOptimizer(params.lr, params.momentum)
tower_grads, tower_loss = [], []
with tf.variable_scope(tf.get_variable_scope()):
images = avatar.batch_data()
images = tf.cast(tf.reshape(images, [-1, 28, 28, 1]),
dtype=tf.float64)
lables = avatar.batch_lable()
conv = dfxp.Conv2d_q(name='conv',
bits=params.bits,
training=False,
ksize=[5, 5, 1, 20],
strides=[1, 1, 1, 1],
padding='VALID')
self.conv_in = images
self.batch_in, self.conv_w, self.x_s, self.w_s, self.conv_w_q = conv.forward(
self.conv_in)
batch = dfxp.Normalization_q(name='batch',
bits=params.bits,
num_features=20,
training=True)
self.scale_in = batch.forward(self.batch_in)
scale = dfxp.Rescale_q(name='scale',
bits=params.bits,
training=False,
num_features=20)
self.relu_in, self.scale_w, self.scale_b = scale.forward(
self.scale_in)
relu = dfxp.ReLU_q()
self.pool_in = relu.forward(self.relu_in)
pool = dfxp.MaxPool_q(ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
flat = dfxp.Flatten_q(12 * 12 * 20)
self.fc1_in = pool.forward(self.pool_in)
self.flat = flat.forward(self.fc1_in)
fc1 = dfxp.Dense_q(name='dense1',
bits=params.bits,
training=False,
in_units=12 * 12 * 20,
units=100)
self.fc2_in, self.w1, self.b1 = fc1.forward(self.flat)
fc2 = dfxp.Dense_q(name='dense2',
bits=params.bits,
training=False,
in_units=100,
units=10)
self.softmax_in, self.w2, self.b2 = fc2.forward(self.fc2_in)
self.loss = tf.nn.softmax_cross_entropy_with_logits(
labels=lables, logits=self.softmax_in)
self.conv_indiff = tf.gradients(self.loss, self.conv_in)
self.batch_indiff = tf.gradients(self.loss, self.batch_in)
self.scale_indiff = tf.gradients(self.loss, self.scale_in)
self.train_step = optimizer.minimize(self.loss)
self.init_op = tf.global_variables_initializer()
self.saver = tf.train.Saver()
self.summary = tf.summary.merge_all()
self.graph.finalize()